Price and Network Dynamics in the European Carbon Market

HAL Id: halshs-01484117https://halshs.archives-ouvertes.fr/halshs-01484117

Submitted on 6 Mar 2017

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.

Price and Network Dynamics in the European CarbonMarket

Andreas Karpf, Antoine Mandel, Stefano Battiston

To cite this version:Andreas Karpf, Antoine Mandel, Stefano Battiston. Price and Network Dynamics in the EuropeanCarbon Market. 2017. �halshs-01484117�

https://halshs.archives-ouvertes.fr/halshs-01484117

https://hal.archives-ouvertes.fr

Documents de Travail du Centre d’Economie de la Sorbonne

Price and Network Dynamics in the European

Carbon Market

Andreas KARPF, Antoine MANDEL, Stefano BATTISTON

2017.10

Maison des Sciences Économiques, 106-112 boulevard de L'Hôpital, 75647 Paris Cedex 13 http://centredeconomiesorbonne.univ-paris1.fr/

ISSN : 1955-611X

Price and Network Dynamics in the European

Carbon Market

Andreas Karpf§, Antoine Mandel¶, Stefano Battiston‖

Abstract

This paper presents an analysis of the European Emission TradingSystem as a transaction network. It is shown that, given the lack of acentralized market place, industrial actors had to resort to local con-nections and financial intermediaries to participate in the market. Thisgave rise to a hierarchical structure in the transaction network. To em-pirically relate networks statistics to market outcomes a PLS-PM mod-eling technique is introduced. It is shown that the asymmetries in thenetwork induced market inefficiencies (e.g increased bid-ask spread).Albeit the efficiency of the market has improved from the beginningof Phase II, the asymmetry persists, imposing unnecessary additionalcosts on agents and reducing the effectiveness of the market as a miti-gation instrument.

1 Introduction

The European Union Emission Trading Scheme (EU-ETS) is the cor-nerstone of European climate policy. On the one hand, it should allowEurope to reduce its carbon emissions at the least possible cost (see e.gStavins, 1995). On the other hand, it should induce economic actorsto account for the cost of carbon in their investment decisions (see e.gLaing et al., 2013; Koch et al., 2014). To fulfill these objectives, theprice of carbon has to be a strong and stable signal, the carbon markethas to aggregate information efficiently and rapidly.

The history of the ten first years of the market, summarized infigure 1, shows it has failed dramatically in both respects. Prices havebeen extremely volatile, participation has been restricted, informationhas been aggregated slowly and inefficiently. A characteristic failure isthe fact that the massive overallocation of allowances at the beginningof phase I was diagnosed only after the first reporting period and notendogenously by the market.

§corresponding author,Centre d’economie de la Sorbonne, University Paris 1 Pantheon-Sorbonne

¶Centre d’economie de la Sorbonne,Paris School of Economics, University Paris 1Pantheon-Sorbonne

‖Department of Banking and Finance, University of Zurich

1

Documents de travail du Centre d'Economie de la Sorbonne - 2017.10

These symptoms of inefficiency have been clearly established in theliterature. We also argue that the root cause of this inefficiency caneasily be grasped by intuition : a poor market design that hadn’t fore-seen the need to create a centralized exchange platform. Yet, we alsoargue that there are still lessons to be learnt by looking at the mechan-ics of failure. A unique feature of the European carbon market is theavailability of a data set which contains all the transactions performedon the market. Therefrom, the complete transaction network can bereconstructed. One can then relate the evolution of the structure ofthe network to the emergence of market inefficiencies.

We hence follow the growing strand of literature that investigatesmarket dynamics with a network-based approach, to gain a detailedunderstanding of the structure of the EU-ETS market and the rela-tionships between network structure, informational asymmetries andmarket dynamics. Therefore a set of empirical relationships betweenthe structure of the trade network and the outcome/efficiency of themarket is established. More specifically with regard to the latter, wetrack the evolution of prices, bid-ask spreads and profits. These empiri-cal relationships can be used to track future developments in Europeancarbon trading but also to assess the efficiency of other markets.

Our analysis shows that in the absence of a central market place,agents had to resort to local networks and financial intermediaries toexchange emission certificates. This led to the emergence of hierarchi-cal and assortative networks with fat tailed degree distributions, whichturned out to be rather inefficient in terms of the price discovery mech-anism and the incorporation of new information. It is demonstratedhow the hierarchy in the market as implied by the observed fat taileddegree distributions as well as the hierarchical structure of the net-work can be associated to informational asymmetries in the market.We further show how informed traders can be characterized in termsof centrality measures, and how the evolution of connectivity patternscan serve as an indicator for volatility or liquidity on the market. Wefind that market efficiency improved during Phase II as the share ofspot market trading increases. It is however also shown that the majorflaws of the EU ETS in principle persist.

The paper also provides a methodological contribution by intro-ducing a Partial Least Squares Path Modeling (PLS–PM) approach todefine endogenously the temporal evolution of the network rather thanresorting to an exogenously fixed time-window. Using this approachallows to investigate the structural evolution of the trading network ina dynamic manner.

The remainder of the paper is organized as follows. Section 2 re-views the related literature. Section 3 provides a description of the or-ganization, the history and the data of the European Emission TradingSystem. Section 4 and 5 provide respectively a static and a dynamicanalysis of the network. Section 6 concludes.

2


start of Pilot Phase I

first release of verified emission data start of Phase II, integration of CDM/JI

approx. begin of VAT fraud

Lehman Brothers collpase

decision to include airline industry

news about VAT fraud emerge

trading halt after cyber attack

0

10

20

30

2005 2006 2007 2008 2009 2010 2011Date

EU

A p

rice

(EU

R)

Figure 1: The price of the EUAs over time, annotated with important marketevents

2 Related literature

The EU-ETS being the first large-scale carbon market in existence,its performance has been extensively analyzed in the literature. Acomprehensive overview of the design, the history and of the earlyliterature on the EU-ETS is given in (Ellerman, 2010). The existenceof a very detailed dataset about the market has also given rise to anumber of firm-level analysis, in particular with regard to determinantsof market participation (see in particular Jaraite et al., 2012; Zaklan,2013; Koch et al., 2014). However, to our knowledge, this paper is thefirst to focus on transaction-level data.

More specifically related to this paper, a first strand of literaturehas focused on the determinants of the price formation process on themarket. With respect to the determinant of prices, most empiricalstudies (see e.g Hintermann, 2010; Creti et al., 2012; Koch et al., 2014)find that fundamentals such as economic activity and energy prices,have significant explanatory power for EUA prices. Energy prices areinsofar of great importance, as the energy sector is the dominant sectorin the EU ETS, and fuel-switching represents a relatively cheap andeffective method for power plants to reduce the greenhouse gas emis-sions. However, Hintermann (2010) also emphasizes, that, especiallyin the beginning of Phase I, EUA prices were not driven by abatement

3


costs and that new information, with regard to fundamentals, like elec-tricity prices or weather, were only incorporated with a substantial lag,thus indicating inefficiencies in the price formation process. Creti et al.(2012) emphasize a similar discrepancy between phase I and phase IIby putting forward the presence of different cointegration relationshipsbetween the phases. They also put forward that the influence of “fun-damentals” materialize more clearly in the second phase and identifya number of complementary price drivers such as policy events. Kochet al. (2014) provide similar quantitative results and additionally findthat the usage of renewables has a significant explanatory power onthe price.

With respect to the efficiency of the market, existing results arerather negative. Palao and Pardo (2012) show the presence of priceclustering in the European Carbon Futures Markets, which they in-terpret as a sign of inefficiency and as a weakness of the price signalprovided by the market. Charles et al. (2013) investigate the spot andfuture carbon markets with regard to cost-of-carry hypothesis. Thecost-of-carry model is rejected in all cases implying arbitrage possibili-ties and thus an inefficient functioning of the market. High transactioncosts have also been identified as an important source of inefficiency(see e.g Jaraite et al., 2012; Zaklan, 2013). They have led to restrictedparticipation and hence reduced the scope and the efficiency of themarket . Martin et al. (2014) even points out participants have of-ten perceived the EU–ETS as a compliance mechanism rather than amarket-based policy. There exist relatively few theoretical analysis ofthe micro-economic sources of inefficiency on the carbon market. Med-ina et al. (2014) put forward the role of informational asymmetries asan important source of trading frictions and show that these frictiondecreased from phase I to phase II.

This paper provides a complementary, more structural, perspec-tive on the issue through a network-based analysis. In this respect,our contribution relates to the burgeoning stream of literature, whichemphasizes the analysis of the topological properties of transactionnetworks as a way to capture in a condensed manner the structuralproperties of markets. These approaches where applied in a varietyof fields ranging from micro-finance/economics (Banerjee et al., 2013;Vignes and Etienne, 2011) through international trade (Duenas andFagiolo, 2014) to financial markets (Battiston et al., 2012a,b; Chinazziet al., 2013; Iori et al., 2008).

3 The European Emission Trading System

The European Emission Trading System (ETS) is used to implementEuropean decisions on emission reductions in a range of industrial sec-tors, the so-called ETS sectors, which correspond to close to 50% of thegreenhouse gas (GHG) emissions in Europe. More precisely, the ETScovers, above certain capacity thresholds, power stations and othercombustion plants, oil refineries, coke ovens, iron and steel plants andfactories making cement, glass, lime, bricks, ceramics, pulp, paper and

4


board as well as, since 2012, aviation.Emission reductions in non-ETS sectors (mainly transport, agri-

culture and waste) are enforced by a range of regulatory measures.Emission reductions in the ETS sector are enforced by a unique andfully market-based mechanism based on the “cap-and-trade” principle.Namely, a certain amount of allowances (EUAs or European AllowanceUnits in the context of the ETS) are allocated or auctioned to emit-ters of greenhouse gases, e.g on the basis of historical emissions. Theseemitters also have to surrender each year an amount of allowance corre-sponding to their actual emissions, which are permanently recorded attheir installations. Allowances therefore become a scarce resource. Asthey are moreover legally tradable, there de facto exists a market forallowances, usually referred to as the carbon market. On this market,emitters which have reduced their emissions below their quota can selltheir remaining allowances, while emitters which have exceeded theirquota can buy available allowances (and hence avoid the payment ofa fine). The system is supposed to exploit differing marginal abate-ment costs (MAC), i.e. the marginal cost of reducing green house gasemissions by one unit, between countries, firms, industries or even be-tween different branches within a company (see Ellerman and Decaux,1998). Emitters from high technology sectors with relatively low levelsof emission could thereby profit from the comparatively low abate-ments costs of other companies, e.g from emission intensive industries.In theory, agents should then arbitrate between emission trading andinvestments in emission reduction technology until marginal abatementcosts are equalized. In theory the aggregate emissions should therebybe reduced at the least possible cost.

The actual setting of the ETS is slightly more complex than thistheoretical template. On the supply side, on top of the EUAs issued bymember states, allowances can be generated through the clean develop-ment (CDM) and joint implementation (JI) mechanisms, which certifyemission reductions performed respectively in developing and transi-tion economies. On the demand side, some organizations purchase al-lowances in order to discard them or to compensate other emissions andhence reduce the aggregate emission ceiling. Additionally, financial in-stitutions can enter the market as brokers or for other “non-industrial”motivations. In practice, the ETS distinguishes between operator hold-ing accounts (OHA), which correspond to installations/companies witha legal obligation to participate in the ETS, and other actors referredto as private holding accounts (PHA).

From the institutional perspective, the market has been organizedin phases: pilot phase I (2005 - 2007), phase II (2008-2012), and phaseIII (2013 - 2020). Since phase II operators are allowed to bank emissioncertificates, this means to use a potential surplus of allowances in thenext trading period. Borrowing is insofar possible as the allocationof allowances takes place in February each year but operators have tosurrender the EUAs only by the end of April (European Commission,2016). There is no centralized trading platform and allowances havebeen traded bilaterally over the counter, via a broker or on one the

5


competing climate exchange markets such as the European ClimateExchange to the European Energy Exchange. The most common formof transactions over the period analyzed in this paper was over thecounter.

In practice, the equalization of marginal costs and the reductionof emission at the least possible costs shall materialize only if thereis full participation in the market and the market settles rapidly atequilibrium. The long-term instability of the price illustrated in Figure1 suggests otherwise. While the price of EAUs in the first year of theEU ETS was nearly constantly above twenty euros per ton of carbondioxide (tCO2), the publication of first verified emission data in April2006 indicated a large overallocation of certificates and resulted intoa market crash which brought EUA prices down to a level close tozero from mid-2007 on. With the beginning of Phase II prices bouncedback to pre-“compliance-break” (Alberola and Chevallier, 2009) levelsin mid-2008. In subsequent years the price of EUAs fluctuated betweenten and fifteen Euros and latter fell close to five euros. Part of thisvolatility can be explained by exogenous drivers such as the financialcrisis or the dynamics of European and Global climate policy (Kochet al., 2014). It is also the case that the history of the EuropeanEmission trading system was accompanied by glitches such as a massiveVAT carousel fraud (Creti et al., 2012) or the fishing attack of January2011, which possibly distorted expectations and the price formationprocess.

Yet, the micro-economic channels through which these exogenousshocks have impacted the market remain unclear. It is also an openquestion whether part of this volatility/inefficiency has emerged en-dogenously from the micro-economic workings of the market. TheETS however presents a unique opportunity to investigate these issuesthanks to the availability of comprehensive transaction data. By law1,every transaction in the ETS has to be recorded in a public registrymaintained by the European commission, the Community Indepen-dent Transaction Log (CITL). The complete dataset is made publiclyaccessible with an embargo of three years2. In order to grasp the re-lationships between these micro-level transactions and the macro-leveldynamics of the market, a natural approach is to analyze the structuralproperties of the trade network and its evolution using the transactiondata. More generally, the creation of the EU ETS is a unique opportu-nity to observe the formation of a market in vivo and thereby increaseunderstanding of market processes in general (Hardle and Kirman,1995).

1directives 2003/87/EC and 2009/29/EC2See http://ec.europa.eu/environment/ets/

6


Figure 2: The ETS network plotted with respect to the geographical position(left) and the centrality (fight) of nodes (yellow and blue indicate financialand industrial agents respectively, the edge color is the same as that of thesource node)

In order to perform this analysis, we extracted the entire transac-tion and compliance data set from 2005 to 2011 using the Python li-brary “scrapy”. This data set comprises 520,000 transactions as well asdetailed information about, installations, accounts and the complianceprocess (allocation, surrendering, verified emissions). For our analysis,transactions connected to the administration of the ETS by govern-ments were discarded3. This brings down the number of transactions toapproximately 364,000. The dataset was augmented by ownership datafrom the “Ownership Links and Enhanced EUTL Dataset” (Jaraiteet al., 2013) and spot prices of the European Allowance Units fromBloomberg. Using the address information contained in the accountinformation we added geocoding to the data set (longitude, latitude)and location specific temperature information. This synthetic datasetdelivers a comprehensive picture of the European emission market.

4 The static trading network

In the remainder of the paper we adopt a network perspective on theEuropean emission market, in which transactions are regarded as di-rected edges between a seller (source vertex) and a buyer (target ver-tex). We first adopt a static perspective where the network is formedby the set of all transactions independently of their time stamp. Fig-ure 2 provides a graphical representation of this aggregate network andunderlines the presence of different groups of agents in the network. In-dustrial actors are the largest group (6384 actors). They correspond to“operator holding accounts”, that is companies which form the primarydemand and supply of allowances as they have a legal obligation to sur-render allowances. The second largest group (2688 agents) consists in

3Namely, only transaction types 3-0, 3-21 and 10-0 were considered

7


CDM

financefinance

foundation

government

industry

foundation

government

industry

AT ATBE

BG

CZ

DE

DK

EE

ES

FI

FR

GB

HU

IE

IT

LI LV

NL

PL

RO SE

SI

SK

other

BE

GR

LU

PT

BG

CZ

DE

LT

DK

EE

ES

FI

FR

GB

GR

HU

IE

IT

LILT

LU

LV

MT MT

NL

PL

PT

ROSE

SI

SK

other

Figure 3: Transaction flows between different groups

financial actors, which enter the market as brokers or for financial engi-neering considerations. The third biggest group (106 agents) are agentswhich participate in the European emission market in the context ofthe Clean Development Mechanism (CDM). Two smaller groups aregovernmental organizations (40) as well as foundations/activists (24)4.The core purpose of the latter group is the reduction of greenhouse gasemissions. Some, not all, try to increase the price of EUAs by buyingand discarding certificates.

4.1 The determinants of network formation

The complex structure of the ETS network is the outcome of a largenumber of micro-level interactions. Accordingly, in view of explainingthe aggregate properties of the market, we first investigate the deter-minants of individual transactions.

4.1.1 Exogenous drivers

We first focus on the exogenous drivers of link formation. In thisrespect, Figure 3 displays the flows of transactions between type andcountry groups. It indicates there is a tendency for agents to tradewith peers from the same country and from different types. In otherwords, the network seems to display strong assortativity with regardsto the location of agents and dissortativity with regards to their type.

4As mentioned above, in our analysis only transactions deemed to be relevant for theprice formation were considered. Transactions connected to the administration (attribu-tion of emission certificates, surrendering etc.) of the market were discarded. Govern-mental organizations thus appear here as normal market participants, who buy and sellcertificates. There role is thus not as dominant as one could expect otherwise. Theirpresence in the analysis is not altering the results.

8


In order to confirm these insights quantitatively, we ran an Expo-nential Random Graph model5 on the EU ETS trading network, seenas an unweighted directed network. Table 1 displays the results of ouranalysis6. It confirms that agents have a propensity to trade with peersfrom the same country (strictly positive and significant coefficient ofthe term nodematch.country). With respect to the mixing behavioramong groups, the results clarify the importance of financial agents.They attract transactions from their industrial peers (the coefficients ofthe terms mix.type.industry.finance and mix.type.finance.industry areboth positive and significant) and also strongly interact among them-selves (homophily measured by coefficient mix.type.finance.finance).Oppositely, industrial agents exhibit a strong propensity to trade withpartners of different type (heterophily).

Table 1: Results of the ERG model estimated on the EU ETS network

EU ETS trading network

edges −8.784∗∗∗

(0.033)transitive 0.376∗∗∗

(0.002)nodematch.country 2.813∗∗∗

(0.012)mix.type.finance.finance 0.944∗∗∗

(0.034)mix.type.industry.finance 0.226∗∗∗

(0.034)mix.type.finance.industry 0.287∗∗∗

(0.034)mix.type.industry.industry −0.998∗∗∗

(0.034)Akaike Inf. Crit. 457,197.300Bayesian Inf. Crit. 457,311.100

Notes: ∗∗∗Significant at the 1 percent level.∗∗Significant at the 5 percent level.∗Significant at the 10 percent level.

These results confirm that, lacking a central market place, indus-trial agents resort to local networks and brokerage services from the

5We used the ergm package of the R statnet library (Handcock et al., 2003)6The edges term can be regarded as a intercept or offset which controls for the degree of

connectedness (density) in the network. The transitive term can be seen as an auxiliaryterm to guarantee robustness of the mixing parameters which are in the center of ourinterest. For the estimation a Maximum Pseudo Likelihood algorithm was used.

9


financial industry in order to trade allowances. The slightly strongereffect observed for transaction from type finance to industry also sug-gests that brokerage services are more intensively used by buyers thanby sellers.

4.1.2 Endogenous drivers

The choice of trading partners does not only depend on exogenousfactors. The importance of the position of an agent in the marketmight also be fundamental in attracting transactions. A network-basedperspective allows to quantify these aspects by relating the formationof links to the existing structure of the network.

The degree is the most basic measure of importance of a node ina network. From a structural perspective, assortativity then measuresthe propensity of nodes to connect to similar nodes in terms of degree.In order to analyze the assortativity of the ETS, we use the Maslov-Sneppen algorithm (see Maslov and Sneppen, 2002), which comparesthe empirical network to a null-model generated by a rewiring pro-cedure7. In the context of the ETS, this null model can be seen asrepresenting the structure that would have expectedly emerge if alltransactions had been performed through the veil of a centralized ex-change platform. The results of the analysis are presented in Figure4 through degree-degree correlation profiles. They show on the onehand a slightly increased connectedness within the group of highlyconnected nodes (the lighter green area in the upper right corner ofthe left hand-side plot) and on the other hand significantly increaseddegree of asymmetric connectedness i.e. between low- and high degreenodes (the yellow to red area along the axes). These results suggestthat large, and hence visible, actors attracted more transactions thanthey would have through a centralized market.

4.2 Degree distribution and Structure of the Trad-ing Network

The network structure that emerges from micro-level interactions canbe analyzed via its degree distribution. In Figure 5 the in-, out- andtotal-degree distributions are displayed on a log-log-scale. The graphclearly shows that the distribution exhibit fat tails. Translated into ourmarket context this means, that there are agents whose degree stronglyexceeds the average. These agents, which are much better connectedthan the average, are likely to play a central role in the functioning ofthe market and hence acquire private information on its operation.

To further investigate the hierarchical structure of the trading net-work, we follow Li and Schurhoff (2014) in analyzing the degree distri-butions in connection with the clustering behavior of agents. Figure 5displays the the in- and out-degrees8 versus the the cliquishness 9 of

7see B8The in- and out-degree, refers to the active and passive connectedness of agents.9Be the k-core of graph a maximal subgraph in which each vertex has at least degree

10


Figure 4: Degree-degree correlation profiles generated by the Maslov-Sneppen algorithm for the ETS as an undirected (left) and directed (right)graph.

agents. In doing so one can check for the presence of clustering andif it is related to the connectedness of agents. For the EU ETS trad-ing network we find a negative relationship between the connectednessof agents (in- and out-degree) and their clustering behavior. In otherwords, the downward sloping cloud implies a hierarchy in the marketwith a strong core of highly connected nodes (to the right on the plot)surrounded by clusters of weaker connected nodes on the outskirts (tothe left of the plot). More precisely, as also illustrated in Figure 2, thetrade network is characterized by a core of strongly connected agentsfrom the finance industry (yellow nodes), surrounded by looser con-nected agents from the industry in the periphery of the graph.

The above described structure is also illustrated by the wave-likeforms of the density of a network statistics (see Figure 7). Financial

k. The cliquishness or coreness of a vertex is then k if it belongs to the k-core but not tothe (k + 1)-core (Seidman, 1983).

11


Figure 5: Degree Distributions

agents are concentrated at crucial location of the network serving ashubs (as measured by their closeness) or intermediaries (as measuredby their betweenness). This structural asymmetry might be a reflectionof the informational asymmetry on the ETS market which authors asMedina et al. (2014) have emphasized.

4.3 Network Structure and Market Efficiency

Our analysis suggests that, in absence of a central market-place, theformation of the EU-ETS trading network was driven by the needfor traders to identify a potential counterpart through geographicalproximity or prominence in the network. These constraints led to theformation of a complex trade network characterized by disassortativemixing behavior (Maslov-Sneppen, see Section 4.1.2), a fat tailed de-gree distributions and a hierarchical structure, with highly connectedfinancial agents in the core and lesser connected industrial agents clus-tering around this core (see Section 4.2).

These asymmetries in the network echo existing evidence in theliterature about market inefficiencies (Hintermann, 2010; Creti et al.,2012; Palao and Pardo, 2012; Charles et al., 2013) and informationalasymmetries (Medina et al., 2014). On the one hand, geographical ho-mophily suggests that differentials in marginal abatement costs wereunderexploited by the market. Indeed, one should otherwise have ob-served large trade flows between eastern and western European coun-

12


Figure 6: Degree vs. Clustering

tries as abatement costs were presumably much lower in the formerthan in the latter. On the other hand, large central agents are likelyto capture information about the functioning of the market and useit to extract informational rents while distorting the price formationprocess. To test this hypothesis, we perform in the following sectionan analysis of the joint evolution of the network structure and marketoutcomes (prices, profits and bid-ask spread).

5 The dynamic trading network

5.1 Overview

An outline of the dynamics of the ETS transaction network is providedin Figure 8, which gives the evolution of key network statistics10 overthe six first years of operation of the market. The picture that emergesis this of a market whose activity and efficiency slowly increases overtime.

The evolution of the network’s degree and density underlines thefact that the intensity of trading in the EU ETS grew over time, witha peak in trade period 5. The sustained high level of betweennesscentrality from period two on indicates that the network has bettertransmission properties, i.e trading in the market is less subject tofrictions. The closeness centrality score, reflecting the average distanceof one node to any node, which reached a peak in trade period four inorder to fall again in successive periods tells a similar story. Degreeassortativity and disassortativity describe the preferential attachmentof nodes to nodes with similar connectivity and vice versa11. While

10see appendix for their definition11At a value of 1 the network exhibits perfect assortative mixing, while a value of −1

implies perfect disassortative mixing. Zero indicates that network is non-assortative.

13


0

3

6

9

0.00 0.25 0.50 0.75 1.00Clustering

Den

sity

typeCDM

finance

foundation

government

industry

0

1

2

3

4

0.00 0.25 0.50 0.75 1.00Closeness

Den

sity

typeCDM

finance

foundation

government

industry

0

1

2

0.00 0.25 0.50 0.75 1.00Degree

Den

sity

typeCDM

finance

foundation

government

industry

0

25

50

75

0.00 0.25 0.50 0.75 1.00Betweenness

Den

sity

typeCDM

finance

foundation

government

industry

Figure 7: Densities plots for various trade network statistics and companytypes in the ETS

the assortativity was steadily decreasing up to trade period three, thefirst period of compliance Phase II, it tends towards zero in subsequentperiods. This implies that agents in the later periods (Phase II) wereless biased in choosing their trading partner and relied to a higherdegree on market signals. This might be due to a higher degree of spotmarket trading in the later trading periods. While trading on spotmarkets only accounted to 40% in 2005, this share rose to 70% in 2007(Convery et al., 2008).

14


degree closeness

transitivity assortativity

betweenness density

0.0

0.2

0.4

0.6

0.8

0e+00

1e−04

2e−04

0.00

0.02

0.04

0.06

−0.15

−0.10

−0.05

0.00

0.00

0.03

0.06

0.09

0.0000

0.0005

0.0010

0.0015

0.0020

2 4 6 2 4 6trade period

Figure 8: Evolution of different network metrics by trade period

5.2 The PLS-PM approach

In the following, we aim to further refine our analysis by linking thedynamics of the network to this of key market indicators such as price,bid-ask spreads and profits. Therefore, a natural approach is to usetime-series of market variables and network-statistics to regress theformer on the latter. However, although market variables have a well-defined time-stamps, the structure of the network at a given date de-pends on the modeler’s assumption about the length of time duringwhich a trading link shall be considered active. A standard approachin the literature (see e.g Panzarasa et al., 2009; Kossinets and Watts,2006; Li and Schurhoff, 2014; Puliga et al., 2014) is to use a slidingwindow on the basis of an assumption about the lifespan of a connec-tion.

15


Degree Pagerank

Closeness Clustering

Katz Assortativity

Betweenness Density

0.00

0.01

0.02

0.000

0.005

0.010

0.00

0.05

0.10

0.15

0.20

0.00

0.02

0.04

0.06

0.08

−0.05

0.00

0.05

0.10

0.000

0.005

0.000

0.005

0.010

0.015

0.001

0.002

0.003

0.004

0.005

2006 2007 2008 2009 2010 2011 2006 2007 2008 2009 2010 2011Date

valu

e

30

90

365

730

1095

1460

Figure 9: Different Network Metrics computed over different window sizes

In the context of the EU-ETS, we lack a natural assumption aboutthe lifespan of connections and, as illustrated in Figure 9, networkcharacteristics exhibit a sizable variability as the window-size increases.In order to overcome this problem and to capture as much as possi-ble of the information provided by the network-structure at differenttime-scales, we use a nested approach via Partial Least Square Pathmodeling (PLS-PM).

The method consists in combining (i) an inner model in which oneseeks to explain dependent variables by latent variables constructedupon network statistics at a multitude of different window-sizes (seeFigure 10)) and (ii) an outer model in which latent network statis-tics are constructed by taking a convex combination of the networkstatistics using different window-sizes in such a way that the explainedvariance of the dependent variables be maximized.

The three inner equations with price, the bid-ask spread and the

16


− 0.8428

− 0.0531

− 0.608

− 0.2943

0.0087

0.29460.0219

− 0.0305

2e−04

− 0.0163

− 0.1438

0.0351

− 0.0126

0.1794

0.0077

7e−04

0.242

− 0.02

− 0.0044

0.0484

0.0063

0.0316

0.0446

− 0.0481

− 0.0296

0.3836

0.0202

0.0245

− 0.6037

− 0.0045

Spread

Profits

Price

other

abatement

AsymmetryDensity

Betweenness

Assortativity

Katz

Closeness

Pagerank

Degree

Figure 10: Inner Model - Latent Network Statistics vs. spread, profits, price

profits as the respective dependent variables are estimated simultane-ously. This delivers unique time series of score values for each networkstatistic which delivers a precise picture of the structural evolution ofEU ETS trading network. The model consists of:

• Eight reflective latent variable equations of the formXjk = λ0jk+λjkLVj + errorjk with the following parameters:12:

– Degree

– PageRank

– Closeness

– Katz

12For a technical description of the different network statistics the reader is referred toAppendix C

17


– Assortativity

– Betweenness

– Asymmetry

– Density

• Two formative latent variable equations of the form LVj = λ0j +λjkXjk + errorj (measurement models) with the following pa-rameters:

– abatement costs

∗ electricity prices (the average of electricity prices fromFrance, Germany, Great Britain and Scandinavia)

∗ oil price (Brent)

∗ temperatures

∗ CO2 commodity price

∗ emissions as recorded in the EU ETS compliance database

∗ country dummy variable

– other

∗ volume traded

∗ the EUROSTOXX stock market indicator as a proxy foroverall economic activity

which, via scores of the form LV j =∑k wjkXjk represent the

embedded manifest variables.

• Three inner model equations (structural equations), explainingrespectively the:

– market price

– bid-ask spread

– profits

A graphical representation of the inner model as well as the es-timated coefficients is displayed in Figure 10. The two formative la-tent variables can be understood as some sort of instrumental vari-able regressions which represent abatement costs (oil price, tempera-tures, CO2 commodity price, emissions, country dummy variable) andother factors (volume traded, the European stock market indicator EU-ROSTOXX as a proxy for economic activity) deemed to be relevant forthe price formation, the bid-ask spread or the ability to generate prof-its out of a trade. The eight reflective latent variable equations eachreflect a different network metric. With the exception of the Asymme-try and the Density variables, the networks are all measured on a nodelevel. The variable Asymmetry corresponds to the Gini coefficient ofthe degree distribution and is a graph level variable like density. Itmeasures the inequality of connectedness among agents and can beinterpreted as a proxy for the Asymmetry in the network. The respec-tive manifest variables in these equations are the respective networkmetrics measured for different window sizes. As mentioned above, ThePLS-PM algorithm now seeks to choose the weights for the manifestvariables in a way, that thereby (via linear combinations) formed la-tent variables, in consideration of the relationships in the inner and

18


outer model, maximize the explained variance of both the latent andobserved dependent variables.

Degree Pagerank Closeness Katz AssortativityBetweenness Density Asymmetry

0.0

0.1

0.2

0.00

0.25

0.50

0.75

1.00

weight

loading

30180 146030180 1460 30180 1460180 146030 180 1460 30180 1460 30180 146030 180 1460Window Size (in days)

Figure 11: Block component weights with respect to different window sizes

Figure 11 displays the weights and the loadings obtained for the dif-ferent network metrics and the different window sizes for which theywere measured (between 1 day and four years). Please note that notall window sizes are preserved in the final model. For the manifestvariables in the respective reflective latent variable equations (all con-tained manifest variables are assumed to reflect different measurementsof the same variable) a positive correlation is prescribed. Individualmeasurements which violated this assumption were removed from theestimation. It is noteworthy that the weighting of window sizes fordifferent network metrics doesn’t behave the same in all cases. Fordegree, PageRank, closeness, clustering and assortativity longer termpositions within the evolving network seem to play a bigger role. Asfar as density (this variable is the only graph-level metric in the estima-

19


tion) and betweenness both long and short term positions seem to bemore relevant than medium term positions. As far as Katz centralityis concerned measurements of window sizes between half a year and ayear seem to be of relatively higher importance.

For the interpretation of the coefficients of the inner model(s) (seeFigure 10), it has to be kept in mind, that, while most network met-rics as well as profits can be computed on a daily and individual basis,density, asymmetry, price and bid-ask spread are only available at theaggregate level. The coefficients for the latter two equations thus de-scribe the influence of the distribution of network metrics, reflectingthe network positions of agents encountering each other on the marketon a certain day, on the respective market price and bid-ask spread.The inner model regressions exhibit a relatively good goodness of fit,with R2 values of 0.6291, 0.5630 and 0.0946 for the price, spread andprofits respectively. This corresponds to a goodness of fit of 0.5723overall.

Table 2: Dependent Variable: Price

Estimate Std. Error t value Pr(>|t|)Intercept -0 0.002 -0 1Degree -0.005 0.004 -1.276 0.202

Pagerank 0.020 0.003 6.733 0Closeness -0.048 0.002 -27.797 0

Katz 0.006 0.002 3.848 0.000Assortativity -0.020 0.002 -9.207 0Betweenness 0.008 0.002 3.218 0.001

Density 0.035 0.002 14.243 0Asymmetry 0.000 0.002 0.101 0.920abatement 0.295 0.002 166.990 0

other -0.608 0.002 -337.170 0

Price As far as the inner model regression for prices is concerned,we observe that all network metric with the exception of degree, close-ness and assortativity exhibit a positive sign. Keeping in mind thatthe EU ETS is considered a directed network, in which a transactionestablishes a link between the seller (source node) and the buyer (tar-get node) we can interpret these results as follows: Elevated averagecentrality measures, represented by PageRank, Katz and betweennesscentrality on a certain trading day imply the presence of dominantagents as far as their network position is concerned exercising elevatedpricing power over the more peripheral agents. A similar interpreta-tion is given by Firgo et al. (2015) in the context of the Austrian retailgasoline market network. Asymmetry as far as the connectedness ofagents is concerned seems to have no significant effect on the prices.

Density can be interpreted as a proxy for overall trading intensity.The positive coefficient of this variable hence implies that in our con-text higher trading intensity drives up prices. The negative coefficientsfor degree (the sum of in- and out-degree) and closeness centrality (to-tal closeness) on the other hand imply decreasing prices as the options

20


of agents to close transactions increase. One has to note however, thatdegree is next the only variable with unsatisfying significance. Thepositive coefficient for the abatement block just implies a positive re-lationship between marginal abatement costs and prices for emissioncertificates. This is in line with theoretical considerations.

Table 3: Dependent Variable: Spread

Estimate Std. Error t value Pr(>|t|)Intercept 0 0.002 0 1Degree 0.025 0.004 6.372 0

Pagerank -0.030 0.003 -9.096 0Closeness 0.032 0.002 16.811 0

Katz -0.004 0.002 -2.481 0.013Assortativity 0.001 0.002 0.280 0.780Betweenness -0.013 0.003 -4.859 0.000

Density -0.016 0.003 -6.094 0Asymmetry 0.022 0.002 12.253 0abatement -0.294 0.002 -153.756 0

other -0.843 0.002 -430.812 0

Bid-Ask Spread The bid-ask spread is, as mentioned before, ofteninterpreted as a measure for informational asymmetry or trading fric-tions in the market: the difference between the buyer and seller priceis assumed to reflect diverging degrees of informedness about the truemarket price (see Medina et al., 2014, in this respect). It is thus notsurprising that the presence of central and thus well informed agentsdecreases the bid-ask spread, as it is implied by negative coefficientsfor PageRank, Katz and betweenness centrality. Density in the contextcan also be interpreted as a proxy for information: the denser (morecomplete) a network is, the more agents have traded in the market inthe past (within the window size) and the better their information is,over which they dispose when concluding a transaction. The negativecoefficient for this variable is in line with this interpretation. Not sur-prisingly we observe that asymmetry in the degree distribution is alsoreflected in the bid-ask spread, which as noted earlier is consideredto be informative about the informational asymmetry in the network.The respective coefficient is positive and highly significant. For degree(sum of in- and out-degree), closeness (here total closeness) and as-sortativity the coefficients are positive and thus inverted compared tothe price equation. As argued above, degree and closeness can be seenas proxy for the options based on past transactions agents have whentrying to complete a transaction implying at the same time a higherpotential for disagreement as far as the price is concerned. Assortativ-ity describes the tendency of agents to connect with other agents whichare similar to them. This means that these agents are less informed asthey only dispose of a limited view on the market, potentially leadingto a greater gap between bid and ask prices. It has to be noted how-ever that assortativity is the sole variable in the regression which isnot significant at a satisfying level. The interpretation of the negative

21


abatement coefficient is a bit more involved. Following the principleidea of emission trading OHAs with higher abatement costs are sup-posed or more likely to rather acquire their missing EUAs necessary tocomply to the emission reduction targets on the market, than to makethe costly investments necessary to bring down emissions by means oftechnologically upgrading their installations. This might lead to an in-creased pressure of accepting asked prices overall reducing the bid-askspread.

Profits The coefficients for the profit equation, as noted above, haveto be interpreted in a different manner as they refer to a dependentvariable which, in contrast to prices and the bid-ask spread, is availableon an individual level. Not surprisingly several variables describing thecentrality of agents exhibit positive slopes and are highly significant.This is in line with the findings in the static network of Section 4.The only exception again are degree and density. The interpretationfor the latter is straight forward as higher trading intensity implieshigher competition and thus lower profit margins. The explanationfor the negative coefficient of the variable degree is however not soclear. Profits as pointed out previously are computed in a cumulativemanner. A higher number of transactions in the past without implyingconcrete centrality within the network might incur a higher possibilityfor losses: Many transactions alone are not sufficient to take a centralposition in the network. It depends on whom an agent is trading with.The positive coefficient for abatement costs is puzzling however. Itcan in our opinion only be explained by the overall relatively low priceof EUAs since the compliance break in 2006 resulting from a massiveoverallocation of emission certificates which, as discussed in Section ??,is still today considered a problem. Of a lesser surprise is the negativeand significant coefficient of the variable asymmetry. Asymmetries inthe network are the result of market inefficiencies that impose highercosts on market participants and thus decreases their profits.

Table 4: Dependent Variable: Profits

Estimate Std. Error t value Pr(>|t|)Intercept -0 0.002 -0 1Degree -0.604 0.006 -109.055 0

Pagerank 0.384 0.005 82.033 0Closeness 0.045 0.003 16.483 0

Katz 0.048 0.003 18.900 0Assortativity 0.242 0.003 71.425 0Betweenness 0.179 0.004 48.247 0

Density -0.144 0.004 -37.341 0Asymmetry -0.030 0.003 -11.835 0abatement 0.009 0.003 3.171 0.002

other -0.053 0.003 -18.857 0

22


5.3 Evolution of the Emission Trading network overtime

The PLS-PM estimation gives rise to two important results: First,by the inner model estimations we obtain easily interpretable resultsindicating the influence of the network metrics and exogenous variableshave on the dependent variables. Second, using the weights estimatedin the outer model of the PLS-PM estimation allows us to construct socalled latent scores. In the PLS-PM they enter the structural equation.Here they are used as some kind of surrogate variables for the networkmetrics measured with the unknown “correct” window size. This allowus to gain insights on how the topology of the emission trading networkevolves and how the position of different subgroups within this networkchanges over time.

23


0500

100015002000

0.000.250.500.75

0.000.250.500.751.00

0.00.20.40.6

0.000.250.500.751.00

0

1

2

0.00.20.40.6

0.000.250.500.75

Volum

eD

egreeP

agerankC

losenessK

atzA

ssortativityBetw

eennessA

symm

etry

2005 2006 2007 2008 2009 2010 2011Date

Typefinance

industry

ets

PLS−PM Scores

0.0

0.2

0.4

0.6

0.0

0.2

0.4

0.00.10.20.30.40.5

0.0

0.1

0.2

0.3

0.00.10.20.30.4

0.0

0.2

0.4

0.6

0.0

0.1

0.2

0.3

Volum

eD

egreeP

agerankC

losenessK

atzA

ssortativityB

etweenness

2005 2006 2007 2008 2009 2010 2011Date

Typefinance

industry

ets

Gini Index of PLS−PM Scores

Figure 12: PLS-PM scores of Network Metrics over time for different com-pany types

24


The conclusions drawn from the different network metrics displayedin Figure 12 are ambiguous. In general one can state that market ef-ficiency has improved over time. As the share of spot market tradingrose, it became easier for agents to aggregate information and to par-ticipate in the European emission market. This interpretation is indi-cated by decreasing average centrality scores (Degree, PageRank andKatz) together with decreasing Gini coefficients for the same variables.In short, a higher market participation rate brought down the averageconnectivity of agents. This effect is the most articulated in trade pe-riod four. There the trading intensity in terms of volume traded wasthe highest, while the average connectivity in the trading network wasat the lowest level.

Looking at the closeness and betweenness centrality scores under-lines this interpretation. The average level of closeness centrality ofagents participating in the EU ETS is sharply increasing over timebeginning in trade period four. This can be viewed as an indicator foran increasing flexibility of agents in choosing their trading partner andthus for an in overall better efficiency of the emission market. This isdespite the Gini index for this metric stays relatively stable over time.The betweenness centrality scores point in a similar direction. Thismetric measures the importance of nodes as intermediaries. It thusis not surprising that this role can mainly be attributed to agents oftype “finance”. Across all six trading periods nodes of this type ex-hibited an in average larger value of betweenness centrality than theircounterparts from the “industry”. The overall increase of betweennesscentrality over time as well as the higher degree of inequality for thismetric for agents of type “industry” in the last three periods can how-ever be attributed to the fact, that this intermediator role was partlytaken over by some of the most central agents of this type. An in-creasing share of spot-market trading replacing OTC, which used tobe the dominant transaction form in the beginning of the EU ETS,can evidently be seen as an explanation for that. This interpretationis supported by the evolution of agents’ mixing behavior over time.While (dis)assortative mixing appears to have played an importantrole in the network formation process in the first three periods, thisvalue is close to zero for both types of agents in the remaining threetrade periods. As for the the evolution of closeness centrality this canbe interpreted as an indicator for a more efficient functioning of themarket. While industrial actors in the first years of the ETS had torely to a higher degree on higher connected intermediaries from thefinancial industry in order to trade, this necessity gradually fell awaywith the emergence of spot market trading. The reason why agents oftype “finance” exhibit a higher degree of disassortativity in the firstthree trading periods can be attributed to their role as intermediaries:According to our prior observation that the trading network exhibits ahierarchical structure (see Section 4.2), EUAs are traded from indus-trial actors in the periphery to financial actors in the center and viceversa.

While above results indicate an overall improvement in market ef-

25


ficiency, a closer look at the differences in scores between the differenttypes of agents reveals that some problematic aspects of the organi-zation of the EU ETS persist. In fact, with the exception of Katzcentrality, we observe an increasing gap towards the latter periods inthe centrality measures (Degree, PageRank, Closeness, Betweenness)between agents of type “finance” and agents of type “industry”. Thisimplies a persistent asymmetry between industrial and financial actorsas far as their network position is concerned. In contrast, the overallinequality in terms of market access between these two types of agents,was thus in fact not decreasing but got more pronounced over time.The asymmetry score itself indicates similar. While the asymmetry inthe market had constantly decreased since mid 2007 this metric is onthe rise again since the beginning of 2010.

6 Conclusion

This paper has analyzed the European Emission trading system froma network perspective. Empirically, the transaction network is charac-terized by clustering behavior of agents, a fat tailed degree distributionand a pronounced hierarchical structure. A quantitative analysis of thedrivers of link formation shows that the emergence of this structure isexplained by the fact that, in absence of a centralized market place,agents had to resort to local networks and financial intermediaries toexchange emission certificates: transactions were driven by the needto identify a potential counterpart through geographical proximity orprominence in the network rather than by complementarity in marginalabatement costs.

The hierarchical structure of the network also suggests that centralagents might have captured informational rents and hence influenceon the price formation process as well as “abnormal” profits. In orderto confirm this intuition we have performed an in-depth analysis ofthe joint evolution of network statistics and of key market characteris-tics (prices, bid-ask spread, profits) using a Partial Least Square-PathModeling technique. The results of this analysis definitively confirmsthe impact of central agents on market dynamics. Over time, the over-all efficiency of the market shows some signs of improvement though astrong asymmetry between industrial and financial actors as well as inoverall remains.

In our opinion, these results reflect a major flaw in the organizationof the European carbon market, namely the lack of a central marketplace. Through the period under consideration, trading was largelyperformed through bilateral OTC transactions and industrial actorshad to resort to local networks or financial intermediaries. Hence, thelack of an institutional structure of the market imposed unnecessarycosts on industrial actors, which often do not possess the resources tocollect market-related information. This undermines the key objectiveof the European Emission Trading System to reduce greenhouse gasemissions at the least possible cost.

Market quality improved over time with the increasing share of

26


spot market trading. However, the major flaws outlined above in prin-ciple still exist. These could be partly solved by the establishment ofa central market place on which emission certificates can be traded.This could improve the quality and effectiveness of emission trading inEurope tremendously.

Acknowledgments

The authors acknowledge the support of the EU FP7 FET projectSIMPOL.

27


References

Alberola, E., Chevallier, J., 2009. European carbon prices and bank-ing restrictions: Evidence from phase i (2005-2007). The EnergyJournal, 51–79.

Banerjee, A., Chandrasekhar, A. G., Duflo, E., Jackson, M. O., 2013.The diffusion of microfinance. Science 341 (6144), 1236498.

Barrat, A., Barthelemy, M., Pastor-Satorras, R., Vespignani, A., 2004.The architecture of complex weighted networks. In: Proceedings ofthe National Academy of Sciences of the United States of America.Vol. 101. National Academy of Sciences, pp. pp. 3747–3752.

Battiston, S., Gatti, D. D., Gallegati, M., Greenwald, B., Stiglitz,J. E., 2012a. Liaisons dangereuses: Increasing connectivity, risksharing, and systemic risk. Journal of Economic Dynamics and Con-trol 36 (8), 1121–1141.

Battiston, S., Puliga, M., Kaushik, R., Tasca, P., Caldarelli, G., 2012b.Debtrank: Too central to fail? financial networks, the fed and sys-temic risk. Scientific reports 2.

Bonacich, P., 1987. Power and centrality: A family of measures. Amer-ican Journal of Sociology, pp. 1170–1182.

Centre, U. R., 2014. JI & CDM projects. [online; 2015-09-29].URL http://www.cdmpipeline.org/

Charles, A., Darne, O., Fouilloux, J., 2013. Market efficiency in theeuropean carbon markets. Energy Policy 60, pp. 785–792.

Chinazzi, M., Fagiolo, G., Reyes, J. A., Schiavo, S., 2013. Post-mortemexamination of the international financial network. Journal of Eco-nomic Dynamics and Control 37 (8), 1692–1713.

Convery, F., Ellerman, D., De Perthuis, C., 2008. The european carbonmarket in action: Lessons from the first trading period. Journal forEuropean Environmental & Planning Law 5 (2), 215–233.

Creti, A., Jouvet, P.-A., Mignon, V., 2012. Carbon price drivers: Phasei versus phase ii equilibrium? Energy Economics 34 (1), 327–334.

Duenas, M., Fagiolo, G., 2014. Global trade imbalances: A networkapproach. Advances in Complex Systems 17 (03n04), 1450014.

Ellerman, A. D., 2010. Pricing carbon: the European Union emissionstrading scheme. Cambridge University Press.

Ellerman, A. D., Decaux, A., 1998. Analysis of post-kyoto co2 emis-sions trading using marginal abatement curves. (working paper).

28


Ellerman, A. D., Joskow, P. L., 2008. The European Union’s emis-sions trading system in perspective. Pew Center on Global ClimateChange Arlington, VA.

European Commission, 2016. EU ETS Handbook. Brussels.URL http://ec.europa.eu/clima/publications/docs/ets_

handbook_en.pdf

European Council, 1993. 94/69/EC: Council Decision of 15 December1993 concerning the conclusion of the United Nations FrameworkConvention on Climate Change .URL http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=

CELEX:31994D0069

European Council, 1998. 2106th Council meeting ENVIRONMENTLuxembourg, 16-17 June 1998.URL http://europa.eu/rapid/press-release_PRES-98-205_

en.pdf

European Parliament, 2003. Directive 2003/87/EC of the EuropeanParliament and the Council of 13 October 2003 establishing a schemefor greenhouse gas emission allowance trading within the Commu-nity and amending Council Directive 96/61/EC.URL http://eur-lex.europa.eu/legal-content/EN/TXT/PDF/

?uri=CELEX:02003L0087-20090625&from=EN

Firgo, M., Pennerstorfer, D., Weiss, C. R., 2015. Centrality and pric-ing in spatially differentiated markets: The case of gasoline. Inter-national Journal of Industrial Organization 40, 81–90.

Freeman, L. C., 1979. Centrality in social networks conceptual clarifi-cation. Social Networks 1 (3), pp. 215–239.

Handcock, M. S., Hunter, D. R., Butts, C. T., Goodreau, S. M., Morris,M., 2003. statnet: Software tools for the Statistical Modeling ofNetwork Data. Seattle, WA.URL http://statnetproject.org

Hardle, W., Kirman, A., 1995. Nonclassical demand: A model-freeexamination of price-quantity relations in the marseille fish market.Journal of Econometrics 67 (1), 227–257.

Hintermann, B., 2010. Allowance price drivers in the first phase of theEU ETS. Journal of Environmental Economics and Management59 (1), pp. 43–56.

Iori, G., De Masi, G., Precup, O. V., Gabbi, G., Caldarelli, G., 2008.A network analysis of the italian overnight money market. Journalof Economic Dynamics and Control 32 (1), 259–278.

Jaraite, J., Jong, T., Kazukauskas, A., Zaklan, A., Zeitlberger, A.,2013. Ownership links and enhanced eutl dataset.

29


Jaraite, J., Kazukauskas, A., et al., 2012. Firm trading behaviour andtransaction costs in the european union’s emission trading system:An empirical assessment. Tech. rep., CERE-the Center for Environ-mental and Resource Economics.

Katz, L., 1953. A new status index derived from sociometric analysis.Psychometrika 18 (1), pp. 39–43.

Koch, N., Fuss, S., Grosjean, G., Edenhofer, O., 2014. Causes of theeu ets price drop: Recession, cdm, renewable policies or a bit ofeverything?—new evidence. Energy Policy 73, 676–685.

Kossinets, G., Watts, D. J., 2006. Empirical analysis of an evolvingsocial network. Science 311 (5757), pp. 88–90.

Laing, T., Sato, M., Grubb, M., Comberti, C., et al., 2013. Assessingthe effectiveness of the eu emissions trading system. Grantham Re-search Institute on Climate Change and the Environment WorkingPaper (106).

Li, D., Schurhoff, N., 2014. Dealer networks.

Martin, R., Muuls, M., Wagner, U. J., 2014. Trading behavior in theeu emissions trading scheme. Available at SSRN 2362810.

Maslov, S., Sneppen, K., 2002. Specificity and stability in topology ofprotein networks. Science 296 (5569), pp. 910–913.

Medina, V., Pardo, A., Pascual, R., 2014. The timeline of tradingfrictions in the european carbon market. Energy Economics 42, 378–394.

Page, L., Brin, S., Motwani, R., Winograd, T., 1999. The pagerankcitation ranking: bringing order to the web.

Palao, F., Pardo, A., 2012. Assessing price clustering in european car-bon markets. Applied Energy 92, pp. 51–56.

Panzarasa, P., Opsahl, T., Carley, K. M., 2009. Patterns and dynamicsof users’ behavior and interaction: Network analysis of an onlinecommunity. Journal of the American Society for Information Scienceand Technology 60 (5), 911–932.

Puliga, M., Caldarelli, G., Battiston, S., 2014. Credit default swapsnetworks and systemic risk. Scientific reports 4.

Seidman, S. B., 1983. Network structure and minimum degree. Socialnetworks 5 (3), pp. 269–287.

Stavins, R. N., 1995. Transaction costs and tradeable permits. Journalof environmental economics and management 29 (2), 133–148.

30


United Nations, 1992. United Nations Framework on Climate Change.[online; 2015-09-29].URL https://unfccc.int/files/essential_background/

background_publications_htmlpdf/application/pdf/conveng.

pdf

United Nations, 1998. Kyoto protocol to the united nations frameworkconvention on climate change.URL http://unfccc.int/resource/docs/convkp/kpeng.pdf

Vignes, A., Etienne, J.-M., 2011. Price formation on the marseille fishmarket: evidence from a network analysis. Journal of Economic Be-havior & Organization 80 (1), 50–67.

Wasserman, S., Faust, K., 1994. Social network analysis: Methods andapplications. Vol. 8. Cambridge University Press.

Zaklan, A., 2013. Why do emitters trade carbon permits?: Firm-levelevidence from the european emission trading scheme. Tech. rep.,DIW Berlin, German Institute for Economic Research.

31


A EU ETS - the detailed legal and orga-nizational background

A.1 Historical Backkground and Legal Foundations

The Kyoto Protocol (KP) United Nations (1998) from 1998 extendedthe United Nations Framework Convention on Climate Change (UNF-FCCC) United Nations (1992) negotiated in 1992 during the UN Con-ference on Environment and Development by defining targets for thereduction of green house gas (GHG) emissions into the atmosphere.These targets follow the principle of “common but differentiated respon-sibilities” as outlined in Article 3 of the KP United Nations (1998).Accommodating the responsibility of industrialized countries for thecontemporary levels of GHG emissions, these targets were determinedto be binding for the group of developed signatory states referred toas the Annex 1 parties. The protocol was signed and ratified by 191parties of which one was the European Union13.14

The Annex 1 parties comprise 37 industrialized countries of which29 are now members of the European Union. The legally binding com-mitment of the signatory countries concerns the most relevant green-house gases and gas groups.15 The targets themselves are howeverquantified in CO2 equivalents with regard to global warming poten-tial and as percentages of the emissions in a base year, which, for themajority of the Annex 1 parties, is 1990. The European Union as awhole committed itself to collectively reduce CO2 emissions by 8% un-til 2012 and 20% until 2020 with respect to base year emissions. Underthe premise of “common but shared responsibilities” member state spe-cific reduction goals were defined, which take into account the differentlevels of economic development within the union, the respective struc-tures of the national economies as well as early measures to reduceGHG emissions.

To keep the costs of limiting CO2 emissions as small as possiblefor the signatory countries the KP allows for so called “flexible mecha-nisms” which serve as an alternative to traditional approaches like car-bon taxes or compensating measures as reforestation (Art. 3.3) UnitedNations (1998). These mechanisms comprise International EmissionTrading (IET), Clean Development Mechanisms (CDM) (United Na-tions, 1998, Art. 12) and Joint Implementation (JI) (United Nations,1998, Art. 6) and shall be described in detail in the following subsec-tion.

Since Japan rejected all attempts to give the UN the legal instru-ments to enforce the emission reduction commitments in the KP andthe United States withdrew from the protocol in 2001 it became soonclear that the EU had to find an internal solution if it wanted to stick

13Council Decision of 15 December 1993 European Council (1993)14Noteworthy exceptions are the United States which signed but never ratified the KP

and finally withdrew in 2001 and Canada which quit the treaty in 2011.15Carbon dioxide, methane, nitrous oxide, sulphur hexafluoride, hydrofluorocarbons and

perfluorocarbons

32


to the GHG reduction targets to which it committed itself in the KPEllerman (2010). After an understanding was found between mem-ber states to differentiate the GHG reduction targets with regard tothe level of economic development in the form of the “burden shar-ing agreement” (BSA) European Council (1998), the initial resistancewith regard to the implementation of a European emission tradingscheme (ETS) began to crumble. The European Union emission trad-ing scheme was finally legally implemented by directive 2003/87/ECEuropean Parliament (2003)16 and subsequently adopted into nationallaws.

A.2 The Functioning of the European Emission trad-ing System

Emission Trading The concept of IET plays the central role offlexible emission reduction instruments and is the cornerstone of Euro-pean climate policy. The system bases on a “cap-and-trade” principlein which permitted emission units, so called allowance units are allo-cated to emitters of green house gases. One emission allowance unittypically corresponds to one metric ton of CO2-equivalent. These as-signed allowance units (AAU or EAU European Allowance Unit in thecontext of the EU ETS) normally depend on historical yearly greenhouse gas emission data and are capped with regard to committednational or supranational emission reduction targets. The allocatedallowance units then correspond to the amount of emissions an op-erator is allowed to produce within the upcoming year. Allowanceunits thereby become a scarce good, which participants can exchangein a market context. Periodically market participants which are legallycommitted to reduce their emissions have to surrender the amount ofallowance units in their possession. These are subsequently comparedwith the realized emissions which are permanently recorded at the re-spective installations to check if the emission reduction targets weremet. If the available allowance units fall short of the realized emis-sions, the obliged market participants have to pay a fine proportionalto the allowance units by which the emission reduction obligationswere missed. Emitters therefore face a choice between reducing theiremissions by targeted investments or by acquiring additional emissioncertificates on the market to meet legal obligations. The system ideallyis supposed to exploit differing marginal abatement costs (MAC), i.e.the marginal cost of reducing green house gas emissions by one unit,between countries, firms, industries or even between different brancheswithin a company Ellerman and Decaux (1998). Emitters from hightechnology sectors with relatively low levels of emission could therebyprofit from the comparatively low abatements costs of other companiesfrom emission intensive industries and vice versa. Installations can befactories, power plants or even aircrafts.

16The directive was later amended by Directive 2004/101/EC, Directive 2008/101/EC,Regulation (EC) No 219/2009 and Directive 2009/29/EC

33


Clean Development Mechanism & Joint ImplementationThe system of emission trading is complimented by the CDM and theJI mechanism. In contrast to emission trading these mechanisms areproject based. Predicated on the assumption that actions which lead tothe reduction of GHG eventually have positive effects in slowing downglobal warming no matter where on the planet they are conducted, An-nex 1 countries can engage in GHG emission reducing projects abroadin order to earn emission reduction units (ERU) which in turn canbe traded on the emission market or used when surrendering one’s al-lowances at the end of a compliance period. While the JI mechanismis supposed to foster cooperation between Annex 1 countries17 to meettheir GHG reduction targets, the CDM aims to stimulate GHG reduc-ing investments and projects in non-Annex 1 countries (mainly devel-oping countries) to promote sustainable development (United Nations,1998, Art. 12) and to help Annex 1 countries to meet their emissionreduction commitments with the lowest possible costs.

The implementation of IET in the European contextThe ETS covers factories, power stations, and other installations witha net heat excess of 20 MW in emission intensive industries responsi-ble for roughly 50% of the GHG emissions in the concerned 31 coun-tries (EU plus Switzerland, Norway and Liechtenstein). With directive2008/101/EC the aviation industry was also included into the ETS.The emission allowance units (EUA) are allocated to each of the ap-proximately 11,000 installations in February each year on a nationallevel in line with the respective BSA and KP reduction targets andhave to be surrendered by the operator holding accounts (OHA) atlatest end of April in the subsequent year. The fine for each EUAafter surrendering that falls short of the verified emissions amountsto EUR 100. Operators are allowed to bank and respectively borrowallowances within a trading period. It was however not permitted tocarry allowances from Pilot Phase I (2005 - 2007) to Phase II (2008-20012), and from there to Phase III (2013 - 2020) Ellerman and Joskow(2008). The ETS is not only open to OHAs, but also private entitieswhich don’t fall under the ETS regulation are allowed against a feeto trade on the emission market. These entities are referred to as pri-vate holding accounts (PHA). EUAs can be traded bilaterally, over thecounter (OTC) via a broker or on one of Europe’s climate exchangemarkets like the European Climate Exchange (ECX), the EuropeanEnergy Exchange AG (EEX) etc. For the time for which the transac-tion data set is available the most common form of transactions wasOTC.

17The majority of currently ongoing Joint Implementation projects are situated in tran-sition economies with Annex 1 obligations like the Russian Federation and Ukraine Centre(2014)

34


A.3 Transaction data from Community Indepen-dent Transaction Log

An attractive feature for research on the European Emission marketis the public availability of the entire transaction data set. As pre-scribed by 2003/87/EC and 2009/29/EC every transaction in the ETShas to be recorded in some sort of accounting system (registries) andis accessible to the public with an embargo of three years. At thebeginning these registries were organized on a national level. Since2008 this function is resumed by a central Community IndependentTransaction Log (CITL) accessible online under http://ec.europa.

eu/environment/ets/. On this website of the (CITL) all transac-tions since the beginning of emission trading in Europe in 2005 arepublished. The transaction data from the CITL form the base of ournetwork-based analysis of the EU ETS.

B Maslov-Sneppen random rewiring

A basic method within the class of approaches to investigate the forma-tion process of a network is the Maslov-Sneppen (Maslov and Sneppen,2002) algorithm: comparing the empirical network with a quantity ofrandom networks with an identical degree sequence and distributionallows us to generate degree-degree correlation profiles which permit toidentify connectivity patterns between nodes of different degrees. Theso called null-model is generated by systematically rewiring the origi-nal network: Two pairs of connected nodes A− > B and C− > D arerandomly selected from a network and rewired in the fashion A− > Dand C− > B. If the thereby generated new connections already ex-ist the procedure is aborted and two new pairs of connected nodesare randomly selected and the rewiring attempt is repeated. Doingthis sufficiently often, a rule of thumb suggests a number as high asten times the number of edges, one obtains a random graph with thesame degree sequence and distribution as the original graph. This pro-cedure is repeated multiple times. Then the generated null-modelsare compared with the original network. More precisely, we com-pare the number of edges between two nodes with degrees K1 and K2

in the empirical network N(K1,K2) and the mean in the generatedrandom networks Nr(K1,K2):R(K1,K2) = N(K1,K2)/Nr(K1,K2).If the deviance of the empirical network from the null-model is sig-nificant can be assessed by computing the Z-scores: Z(K1,K2) =(N(K1,K2) − Nr(K1,K2))/sigma(K1,K2), where sigma(K1,K2) isthe standard deviation of Nr(K1,K2).

C Network statistics

1. Asymmetry : The Gini-coefficient of the degree distribution mea-suring the inequality or asymmetry of the connectivity of agents.

2. Degree centrality : Degree centrality measures the connectedness

35


of a vertex. In a network with n vertices a vertex can have atmaximum a degree of n−1 (i.e. links to other nodes). The degreecentrality measure is then the relation of the degree of one nodei to the number of all other nodes: Cedi = di(g)/(n − 1). Thismeans the higher the centrality of one vertex with respect to itsdegree is, the closer to one this centrality measure will be. In thecontext of this study the graph-level degree centrality is used.The score is normalized by the maximum theoretical score for agraph with the same number of vertices.

3. Eigenvector centrality (Bonacich, 1987): This measure considersnot only the connectedness of a vertex but also the connectednessof its neighbors. The centrality therefore not only depends onthe links a vertex has but also on how central its neighbors arewithin the network. If A is the adjacency matrix of the networkN where its elements ai,j ∈ {0, 1} indicate the presence of a link(0 no link; 1 link) between two vertices i and j, and M(i) isthe set of neighbors of vertex i, the eigenvector centrality of avertex is the sum of the centralities of its neighbors multiplied bya constant 1

λ : Cei = 1λ

∑j∈M(i) C

ej = 1

λ

∑i∈N ai,jC

ej . Rearranged

in matrix form one gets the eigenvector equation Ax = λx whichis eponymous for this centrality measure. In the context of thisstudy the graph-level eigenvector centrality is used. The score isnormalized by the maximum theoretical score for a graph withthe same number of vertices.

4. PageRank centrality (Page et al., 1999): This centrality measureis a variety of the eigenvector centrality and represents the likeli-hood that an agent randomly selecting different links will arriveat any particular vertex.

5. Closeness centrality (Freeman, 1979): This centrality measure isdefined as the inverse of the average shortest distance from onevertex to all other vertices. Thus if l(i, j) is the number of linkson the shortest path between the vertices i and j, the averagedistance (number of links) is d(i, j) =

∑i6=j l(i, j)/(n − 1) and

the closeness centrality measure is Ceci = (n− 1)/∑i6=j l(i, j) =∑

i6=j 1/d(i, j). In the context of this study the graph-level close-ness centrality is used. The score is normalized by the maximumtheoretical score for a graph with the same number of vertices.

6. Betweenness centrality (Freeman, 1979): In contrast to the de-gree and closeness centrality the betweenness centrality doesn’tmeasure the centrality of a vertex by the connectedness of a ver-tex but rather by its role as an intermediator. The betweennesscentrality of a vertex i is thus the number of shortest paths be-tween a pair of vertices j and k on which one can find vertexi relative to the number of all shortest paths between j and ksummed over all pairs of vertices. If pjk is the total number ofpaths between two vertices j and k and pjk(i) is the number ofpaths between these two vertices passing through vertex i, be-tweenness centrality is defined as: Cebi =

∑i6=j 6=k pjk(i)/pjk. In

36


the context of this study the graph-level betweenness centrality isused. The score is normalized by the maximum theoretical scorefor a graph with the same number of vertices.

7. Clustering : This metric is sometimes also denoted as the clus-tering coefficient as it measures the probability that the adjacentvertices of a vertex are connected. In the node-level case thedefinition by Barrat et al. (2004) for the transitivity metric isCwi = 1/si(ki − 1)

∑i,h(wij + wij)/2aijaihajh, where si is the

strength of a vertex, which is defined as the sum of edges of alladjacent vertices, ki is the degree of the vertex, aij are elementsof the adjacency matrix and wij are the weights.

8. Katz Centrality (Katz, 1953): Katz centrality computes the rela-tive centrality of a node within a network by measuring the num-ber of neighbors and all other nodes in the network that connectto this respective node via its neighbors.

9. Density : Following Wasserman and Faust (1994) the density ofa graph is defined as the ratio of the number of edges and thenumber of theoretically possible edges.

10. Power : The exponent of the power law distribution fitted to thedegree distribution.

11. Assortativity : The preferential attachment with regard to type(industry, finance etc.), origin and connectedness. Be eij thefraction of edges in a network which connect vertices of type iand j. For undirected networks this value is symmetric: eij = eji.If ai and bi are the fractions for each type of target originatingfrom a vertex of type i, this implies that

∑ij eij = 1,

∑j eij = ai

and∑i eij = bj . The assortative mixing score is then (

∑i eii −∑

i aib1)/(1−∑i aibi).

37
